Technology

Two contexts, two architectures, one rule.

Processing happens where the conversation happens. In the office: inside the building. At home: on the device.

Block diagram

Where recognition runs.

Neither path leaves the space the conversation is happening in.

Counter station

Organisations: municipalities, healthcare, banks, post offices, pharmacies.

Perimeter: the building

  1. Station microphonesAudio captured at the desk, on both sides.
  2. Thin client8-inch display and 5-inch operator module.
  3. Building local networkNo route to the outside.
  4. On-premise server, NVIDIA CUDA platformRecognition, entity extraction, translation and summary.

All inference happens inside the building. No audio, video or text leaves it. No internet connection required.

Home device

Personal use, portable.

Perimeter: the device

  1. Integrated microphonesPicked up in the home.
  2. Edge AI, Qualcomm DragonWing platformModels run on board, on the neural processor: recognition and speaker segmentation.
  3. Screen and RGB LEDsText with one colour per voice, alerts for environmental sounds.

Audio never leaves the device. Integrated battery, no account, no network required.

The software

A thin client, a server that does the work.

The client captures and displays. It does not decide, does not store, and does not talk to anything outside.

  1. Two clients, two implementations

    At the counter the client runs on ESP32, inside the two displays. At home it is a Linux application on the device itself. In both cases it does the same three things: capture audio and video (the video is what sign-language recognition needs), show the text, and accept a tap.

  2. Where there is a server, the models can be bigger

    The counter station has the on-premise server behind it, and that makes room for larger recognition models: fewer errors on the words, and punctuation in the right places. Punctuation is what makes a text readable instead of something to decipher.

  3. At home, everything on board

    On the home device the models are leaner and faster: the edge AI hardware runs recognition, speaker segmentation and, when needed, environmental sound recognition. Fewer parameters, lower latency, no server to reach.

  4. More than one model

    Recognition runs alongside speaker segmentation, entity extraction — addresses, phone numbers, amounts, dates — and translation. That is where the end-of-conversation summary comes from.

  5. Under evaluation: Auracast

    For museums, conference rooms and lecture halls we are evaluating Auracast (Bluetooth LE Audio): a single broadcast that anyone can pick up on a compatible hearing aid or on headphones — no pairing, and nobody has to ask for help.

Auracast™ is a technology evaluation, not an available feature: it is not part of what we install today. Auracast™ is a trademark of Bluetooth SIG, Inc.

Adaptation

Models tuned to the context.

The architecture allows models to be trained on specific terminology domains, adapting recognition to the vocabulary that recurs in the installation context.

In practice: the words that come up every day in that place — forms, procedures, internal names — are recognised better than a general-purpose model would manage.

We do not publish the compute board model or the maximum number of stations one server supports: those figures need a benchmark, not an estimate.

Practical consequences

What a local architecture gets you.

  1. No connectivity dependency

    The service does not stop when the line goes down, and it needs no guaranteed bandwidth on site.

  2. No third-party sharing

    There is no external provider receiving audio or text, so there is no contract to take on trust.

  3. Predictable latency

    The path from audio to text stays inside the local network, with no round trip to a remote service.

  4. Expandable ecosystem

    Stations and devices are added as needs grow, without rebuilding the installation.